From: tag@nrlmry.navy.mil
Date: Fri, 08 Jul 94 16:32:21 pst

     Name: Paul M. Tag
     Organization: Naval Research Laboratory 
     Interests: AI Applications to Navy Needs
     
     NRL Monterey conducts meteorological research for the Navy and has 
     been involved in AI since about 1987.  Our efforts have included 
     expert system development for overwater fog prediction (see Peak and 
     Tag, Monthly Weather Review, 1989, 2641-2653), tropical cyclone 
     movement (Buck Sampson), satellite imagery interpretation (see below), 
     and clear air turbulence forecasting (Jim Peak); next year we plan to 
     develop an expert system addressing Navy weather concerns in the 
     Mediterranean (Richard Bankert and Arunas Kuciauskas).
     
     Neural network use here has addressed satellite imagery interpretation 
     (see below) and cloud classification from satellite imagery (see 
     Bankert, Journal of Applied Meteorology, 1994 (just coming out)).  Of 
     particular interest may be Rich Bankert's work with a Probabilistic 
     Neural Network (PNN).  The PNN is a NN architecture for direct 
     solution of the Bayesian classification rules.  It allows for quick 
     training without sacrifice in accuracy when compared to more common 
     feed-forward back-propagation networks.
     
     Jim Peak has worked on a number of components of our satellite imagery 
     interpretation project and has linked them together into a simple 
     working prototype.  Our goal is to completely automate the 
     meteorological interpretation of cloud patterns visible in satellite 
     imagery.  We have put the meteorological expertise to do the 
     interpretation into an expert system, but the pattern recognition, 
     etc. is the hard part.  We have addressed the segmentation problem 
     with a new NN-based approach.  Rich Bankert has developed a routine to 
     identify segmented cloud patterns (fronts, cells, jet cirrus, etc.) 
     using shape, size, texture, and spectral features within a neural 
     network.  One thing he is lacking to test it is expertly labeled data. 
     If anyone wants to help in this regard, please contact him.  Feature 
     analysis, following segmentation and identification, is accomplished 
     using empirical orthogonal functions and neural nets.  Our overall 
     design for the imagery interpretation is summarized in two papers: 
     Peak and Tag, Bulletin of the American Meteorological Society, 1992, 
     995-1008 and Journal of Applied Meteorology, 1994, 605-616. 
     
     Recently, I have tried testing machine learning methods to determine 
     their applicability to meteorological problems.  In particular, we are 
     using Quinlan's C4.5 program as applied to fog prediction.  Using the 
     data base we used in developing the fog expert system, we are 
     determining whether C4.5 can generate the prediction rules that we 
     already know.  A similar study involves using buoy data off the coast 
     of California.  If there are meteorological data sets out there in 
     which each case has both related predictors (as many as considered 
     important) and definite outcomes, and if the data set is coastal or 
     open-water oriented, I would like to know about them.  Coming up with 
     good data sets addressing meteorological parameters of interest to the 
     Navy is more difficult than I envisioned.
     
     If anyone needs copies of our papers or if there is any interest in 
     (or help for) our work, our e-mail addresses are as follows:
     
     bankert@nrlmry.navy.mil
     kuciauskas@nrlmry.navy.mil
     peak@nrlmry.navy.mil
     sampson@nrlmry.navy.mil
     tag@nrlmry.navy.mil
     
     Paul M. Tag
     Naval Research Laboratory
     7 Grace Hopper Avenue Stop 2
     Monterey, CA 93943-5502 U.S.A.
     (408) 656-4885
     Fax: (408) 656-4769